A k-segments algorithm for finding principal curves

A k-segments algorithm for finding principal curves

Pattern Recognition Letters 23 (2002) 1009–1017 www.elsevier.com/locate/patrec A k-segments algorithm for finding principal curves €se 1 J.J. Verbeek ...

150KB Sizes 12 Downloads 89 Views

Pattern Recognition Letters 23 (2002) 1009–1017 www.elsevier.com/locate/patrec

A k-segments algorithm for finding principal curves €se 1 J.J. Verbeek *, N. Vlassis, B. Kro Computer Science Institute, University of Amsterdam, Intelligent Autonomous Systems, Kruislaan 403, 1098 SJ Amsterdam, The Netherlands Received 19 December 2000; received in revised form 13 July 2001

Abstract We propose an incremental method to find principal curves. Line segments are fitted and connected to form polygonal lines (PLs). New segments are inserted until a performance criterion is met. Experimental results illustrate the performance of the method compared to other existing approaches.  2002 Elsevier Science B.V. All rights reserved. Keywords: Dimension reduction; Feature extraction; Polygonal line; Principal curve; Unsupervised learning

1. Introduction Principal curves are the non-linear generalization of principal components. They give a summarization of the data in terms of a 1-d space non-linearly embedded in the data space. Intuitively, a principal curve ‘passes through the middle of the (curved) data cloud’. Applications include dimension reduction for feature extraction, visualization and lossy data compression. What concerns the first application, the curve can be used to obtain an ordering of the data by projecting the data to the curve. An example of such a use of principal curves is the ordering of ecological *

Corresponding author. E-mail address: [email protected] (J.J. Verbeek). 1 This research is supported by the Technology Foundations STW, applied science division of NWO and the technology programme of the Ministry of Economic Affairs.

species abundance data (Death, 1999). Applications of the proposed principal curve algorithm can also be envisaged in fitting and recognizing complex shapes in images. Several definitions of principal curves have been proposed in the literature. One of the earliest definitions is based on ‘self-consistency’ (Hastie and Stuetzle, 1989), i.e. the curve should coincide at each position with the expected value of the data projecting to that position. Another approach (Kegl et al., 2000) is to define principal curves of length l as curves of length l that achieve the minimum expected squared distance from points to their projection on the curve. The polygonal line algorithm (PLA) finds principal curves in terms of the latter definition. A more probabilistic approach (Tibshirani, 1992) defines principal curves as curves minimizing a penalized log-likelihood measure. The data are modeled by a mixture of the form:

0167-8655/02/$ - see front matter  2002 Elsevier Science B.V. All rights reserved. PII: S 0 1 6 7 - 8 6 5 5 ( 0 2 ) 0 0 0 3 2 - 6

1010

pðxÞ ¼

J.J. Verbeek et al. / Pattern Recognition Letters 23 (2002) 1009–1017

Z

l

pðxjtÞpðtÞ dt;

ð1Þ

0

where t is a latent variable distributed on an arclength parameterized curve of length l. pðxjtÞ is, for example, a spherical Gaussian modeling the noise located on point t of the curve. In addition to the above-mentioned principal curve algorithms, several other methods can be used to tackle this and closely related problems. The generative topographic mapping (GTM) (Bishop et al., 1998) finds a non-linear curve embedded in the data space. Along this curve a mixture of Gaussian kernels is placed to generate a distribution on the data space. Also some vector quantization techniques can be used to find approximations to principal curves, e.g. self-organizing maps (SOM) (Kohonen, 1995) and growing cell structures (GCS) (Fritzke, 1994). These methods incorporate a topology among the prototype-vectors. The original space is then mapped to a discrete 1-d space instead of a continuous 1-d space. Several problems exist with aforementioned methods. A common feature of these is that they consist of a combination of ‘local models’ that are related by a fixed topology. When the data are concentrated around a highly curved or selfintersecting curve, these methods exhibit poor performance. This is due to the fixed topology among the local models and to bad initialization. Also, often one does not know a priori how many ‘local models’ one needs and one has to make an (educated) guess for the number of local models. On the other hand, the definition of (Hastie and Stuetzle, 1989) explicitly excludes self-intersecting curves as principal curves. In general PLA and Tibshirani’s method also fail due to bad initialization in cases where the data are concentrated around highly curved or self-intersecting curves. See Fig. 1 for an illustration (for PLA, the dashed curve passing through the data is the generating curve). In this paper, we use a probabilistic setting to find principal curves by means of maximizing loglikelihood, resembling the method of (Tibshirani, 1992). We assume the data are corrupted by some noise, therefore we do not require the curve to fit

the data exactly. We use the model (1), where pðtÞ is uniform along the curve and pðxjtÞ is a spherical Gaussian located on point t of the curve, with constant variance over all t. The variance is a smoothing parameter, to be set by the user. We restrict the considered curves to polygonal lines (PLs). Our method avoids many of the poor solutions mentioned above (and shown in Fig. 1) by fitting local linear models without any topological constraints. To enable an efficient search among PLs with a varying number of segments and to avoid local minima, we use an incremental method that considers PLs consisting of an increasing number of segments. Hence, it starts with models that are relatively easy to fit and gradually increases their complexity. Overview of the method. By means of local principal component analysis (PCA) we find local linear models. The local linear models are fitted by a (two-step) iterative update scheme. After fitting, the local linear models are combined to form a PL by means of a heuristic search. Next, if appropriate, a new local linear model is added. The location of the new local linear model is found by means of a global search over the data set. Among all PLs encountered by the algorithm, the PL that maximizes an approximation of the log-likelihood of the data is selected. Overview of the paper. In the next sections we discuss the algorithm in detail. Section 2 is concerned with how we fit the local linear models. In Section 3 we describe the method to find ‘good’ PLs given a set of line segments. In Section 4 the objective function is considered. We end the paper with a discussion.

2. A k-segments algorithm In order to find a ‘good’ PL for the data, we first search for a set of k ‘good’ line segments. In the first subsection we discuss how we extend the k-means algorithm (Gersho and Gray, 1992) to an algorithm for finding k-lines. In Section 2.2 we adapt the k-lines algorithm to find line segments that can be used for PL construction. In the last subsection we describe a method to insert a new segment given k-segments and a data set.

J.J. Verbeek et al. / Pattern Recognition Letters 23 (2002) 1009–1017

1011

Fig. 1. A spiral with some noise. Results for (left to right, top to bottom): GTM, PLA, GCS and SOM.

2.1. Extending k-means to k-lines The generalized Lloyd algorithm or ‘k-means’ is a well-known vector quantization method. To extend this method to k-lines, we start with some definitions. A line s is defined as: s ¼ fsðtÞjt 2 Rg, where sðtÞ ¼ c þ ut. The distance from a point x to a line s is defined as dðx; sÞ ¼ inf ksðtÞ xk: t2R

ð2Þ

Let Xn be a set of n samples from Rd . Define the Voronoi regions (VRs) (Gersho and Gray, 1992) V1 ; . . . ; Vk as Vi ¼ fx 2 Xn ji ¼ arg min dðx; sj Þg: j

ð3Þ

Hence, Vi contains all data points for which the ith line is the closest, see Fig. 2 for an illustration. Analog to k-means, the goal is to find k-lines s1 ; . . . ; sk that minimize the total squared distance of all points to their closest line: k X

X

2

dðx; si Þ :

Fig. 2. An example where the data (open dots) are partitioned in three Voronoi regions for two normal segments (V1 and V2 ) and one (V0 ) ‘zero-length’ segment (black dot).

random orientations and locations of the k-lines. Next, alternate between the following two steps until convergence: 1. Determine the VRs. 2. Replace the lines by the first principal component (PC) of their VR. 2 To prove that this algorithm converges, we make two observations:

ð4Þ

i¼1 x2Vi

To find lines that are local optima of (4) we modify the k-means algorithm slightly: start with

2

More precisely: we first translate the data in the VR to obtain zero mean and then take the line along the first PC of the translated data.

1012

J.J. Verbeek et al. / Pattern Recognition Letters 23 (2002) 1009–1017

• The objective function (4) decreases both in steps 1 and 2. For step 1 this follows from the definition of the VRs. For step 2 this follows from the reconstruction error minimization property of the first PC (Ripley, 1996). • Due to the finite cardinality of Xn , there are only a finite number of distinct (Voronoi) partitions of Xn .

from the centroid of the VR, where r2 is the variance along the first PC. To maintain the convergence property, we check whether (4) is actually decreased by the new segment. If (4) is not decreased, we use the segment that includes all projections to the first PC to obtain guaranteed decrease of (4). 2.3. Incremental k-segments

From these two observations it follows trivially that the algorithm achieves a local optimum for (4) in a finite number of steps. 2.2. From k-lines to k-segments Since we want to construct a PL we actually do not search for lines but for line segments instead. We therefore replace step 2 of the algorithm. Instead of using the first PC, we use a segment of it. For example, we could use the shortest segment of the first PC such that all orthogonal projections to the first PC of the points in the VR are included in the segment. We adjust the definition of the VR straightforwardly by using the distances to the segments, i.e. t is restricted to a bounded interval in (2). Using these segments the value of the objective function is not altered. However, two observations suggested another approach: • The objective function (4) is an easy to compute substitute for the log-likelihood of the data given the segments, see Section 4. The length of the segments is not reflected in (4). As a result relatively long segments are found which result in sub-optimal PLs. • In applications of the algorithm we sometimes observed poor performance. This was due to local optima of (4) which could be avoided by removing the restriction that the segments have to include all projections of the points in the VR to the first PC. To include explicit optimization of the segment length would be computationally very costly. Therefore, we settled for a heuristic that works well in practice. Experimentally, we found that in general good performance is obtained if we use segments of the first PC that are cut off at 3r=2

In most applications we do not know in advance how many segments are ‘optimal’ to model the data. Therefore, it makes sense to try (many) different numbers of segments and use some criterion to select the best number. Since running the algorithm for many different numbers of segments is time consuming, we propose to use an incremental strategy. We start with k ¼ 1 segment and use the search method described in the previous subsections to find optimal segments. Next, we insert a new segment and optimize the k þ 1 segments again. The scheme is repeated until some criterion is met (i.e. a predefined maximum of lines is reached or some performance criterion is met). To determine where to insert the new segment, we first compute for each xi in our data set the decrease of (4) if we would place a zero-length segment on xi . Below we explain why we use this criterion. With a zero-length segment we just mean a point c, hence the distance function (2) reduces in this case to dðx; sÞ ¼ kx ck. The VRs of the zerolength segments are defined as before, by (3), see also Fig. 2. To avoid obtaining segments representing too few points, we only consider those zero-length segments for which the cardinality of the corresponding VR is at least three. Let Vkþ1 be the VR of the zero-length segment maximizing the decrease in (4). Then we insert a segment along the first PC of the data in Vkþ1 which is cut off at 3r=2 on each side of its mean, where r2 is the variance of the data in Vkþ1 when projected orthogonally to the first principal component of the data in Vkþ1 . After the insertion of the new segment, again we are ready to do step-wise optimization of the segments. Note that the zero-length segments are not inserted themselves, but only used to select a good position to insert the new segment.

J.J. Verbeek et al. / Pattern Recognition Letters 23 (2002) 1009–1017

The decrease in (4) due to an insertion of a zerolength segment provides a lower bound on the decrease due to the segment insertion that is actually performed. To see this, let Vkþ1 denote the VR corresponding to the zero-length segment at xi . It is a well-known fact (Gersho and Gray, 1992) that for a given finite set of points S Rd the mean m of those points minimizesPthe squared distance 2 function: m ¼ arg minl2Rd x2S kx lk , hence: P P 2 2 x2S kx mk 6 x2S kx xi k . Since the new segment s includes m, the total square distance to the segment the total square distance to P is at most P 2 2 m i.e. dðx; sÞ 6 x2S x2S kx mk . Hence, we have the lower bound. Note that the search for the minimizing xi can be implemented efficiently as described in Appendix A. 3. Combining segments into a polygonal line Until now we have introduced an incremental method to find a set of line segments. The outset, however, was to find a PL maximizing the loglikelihood of the data. In this section we discuss how to link the segments together to form a PL. Since there are in general many ways to link the segments together, we settle for a greedy strategy. First we generate an initial PL and next we make step-wise improvements on the initial PL. To achieve a fast algorithm to find the PL we do not consider the data in the construction of the PL. We define a fully connected graph G ¼ ðV ; EÞ, where the set of vertices V consists of the 2k end-points of the k-segments. Also, we define a set of edges A E which contains all edges that correspond to the segments. A sequence of edges fðv0 ;v1 Þ;ðv1 ;v2 Þ;...;ðvm 2 ;vm 1 Þ;ðvm 1 ;vm Þg in which all edges are distinct is called a ‘path’ (Wilson, 1972). A path is ‘open’ if v0 6¼ vm . An open path that passes through every vertex in the graph exactly once is called a ‘Hamiltonian path’ (HP). Note that we can consider a HP as a set P E. We want to find the HP P minimizing the total cost of the path, under the constraint: A P E. 3 The cost of a path P is defined as lðP Þ þ kaðP Þ, with 3

There always exists a HP for G that contains A since G is fully connected and A connects only pairs of vertices.

1013

0 6 k 2 R being a parameter to be set by the user. The term lðP Þ denotes the length of the path, defined as the sum of the lengths of the edges in P. The length of an edge e ¼ ðvi ; vj Þ is taken as the Euclidean distance between its vertices: lðeÞ ¼ kvi vj k. The second term, aðP Þ, is a penalty term equal to the sum of the angles between adjacent edges. The parameter k controls the trade-off between preferring short paths and paths that do not contain sharp turns. We introduced the kaðP Þ term to implement a preference for ‘smooth’ (not having sharp turns) curves. The smaller k, the smaller the preference for smooth curves. Especially in cases where the generating curve is self-crossing (cf. Fig. 5) setting k > 0 is important to find the right PL. To construct an initial PL we use the greedy strategy outlined below. We call a HP on a subset of V a sub-HP. We start with the original k-segments as k sub-HPs. At each step we connect two sub-HPs with an edge e. Note that the total cost of a (sub-) HP consisting of two sub-HPs Pi and Pj linked together by an edge e is the sum of the costs of each sub-HP plus lðeÞ plus an angle penalty aðeÞ. Fig. 3 illustrates the angle penalty aðeÞ ¼ a þ b incurred by an edge e ¼ ðvi ; vj Þ that connects two sub-HPs Pi and Pj . We assign to each edge e 2 ðE AÞ cost cðeÞ ¼ lðeÞ þ kaðeÞ. The construction algorithm inserts at each step the edge that minimizes cðeÞ over all edges that connect two sub-paths. Summarizing, the procedure is as follows: 1. Start with k sub-HPs defined by A. 2. While there are at least two sub-HPs. 3. Join those two sub-HPs Pi and Pj ði 6¼ jÞ by edge e 2 ðE AÞ such that e minimizes cðeÞ over all edges connecting two distinct sub-HPs.

Fig. 3. Connecting two sub-paths: the angle penalty of ðvi ; vj Þ is the sum of the angles between an edge and the adjacent edges, i.e. a þ b.

1014

J.J. Verbeek et al. / Pattern Recognition Letters 23 (2002) 1009–1017

Due to greedy nature of the construction discussed above, it might well happen that we can obtain a HP with lower cost by some simple modifications of the initial HP. To find such improvements we use a simple result from graph theory (Lawler et al., 1985): finding the optimal (in terms of cost) HP for k vertices can be expressed as finding the optimal solution for a k þ 1 cities traveling salesman problem (TSP). We use the 2-opt TSP optimization scheme (Lawler et al., 1985) to improve our initial HP.

Fig. 4. Illustration of the distances between a point x a segment s and a point f ðtÞ on s.

2

d? ðs; xÞ =ð2r2 Þ log

4. Objective function

2

log l þ dðs; xÞ =ð2r2 Þ þ c;

ð5Þ

where c is some constant dependent on r. This can be seen as follows:

¼ log

2

expð dk ðf ðtÞ; xÞ Þ dt: t2½0;l

Using the incremental method described above, we obtain a sequence of PLs with an increasing number of segments. As stated in Section 1, we search for the PL that maximizes the log-likelihood of the data. Consider a PL of length l as a continuous arclength parameterized 1-d latent variable t embedded in Rd , where the embedding is given by f : ½0; l ! Rd . For simplicity, we assume a uniform distribution pðtÞ ¼ 1=l on t here for t 2 ½0; l. Furthermore, let pðxjtÞ be a spherical Gaussian distribution with mean point t on the PL and covariance matrix r2 I, where I is the d  d identity matrix. It turns out that, for a latent variable distributing uniformly over a line segment s of length l  r, the negative log-likelihood for a point x can be roughly approximated by

log pðxÞ

Z

Z

ð7Þ Let dk ðs; xÞ ¼ inf t2½0;l fdk ðf ðtÞ; xÞg. Then the last integral can be approximated roughly by c2 exp. 2 2 2 Note that dðs; xÞ ¼ dk ðs; xÞ þ d? ðs; xÞ , this leads to Eq. (5). We neglect the effect of higher density occuring at one side of the PL at places where different nonparallel segments are connected. Hence, we approximate the total log-likelihood of the data as n log l þ

k X X

2

dðsi ; xÞ =ð2r2 Þ;

ð8Þ

i¼1 x2Vi

where l is the total length of the PL. In practice, if we take the objective function as a function of k, the number of segments, the global minimum often is the first local minimum of the objective function. So, a simple stopping strategy could be to keep inserting segments until the objective function (8) reaches its first minimum. It is also possible to keep inserting segments until some limit kmax on k is reached. A variety of other stopping criteria is possible.

pðxjtÞpðtÞ dt t2½0;l

¼ logðlÞ þ c1 log

Z exp t2½0;l

kx f ðtÞk 2r2

2

!

5. Discussion dt:

ð6Þ 2

2

We may write kx f ðtÞk as the sum d? ðf ðtÞ; xÞ þ 2 dk ðf ðtÞ; xÞ , see Fig. 4. Since d? ðf ðtÞ; xÞ ¼ d? ðs; xÞ is constant for different t, we can write the last term in (6) as

To test the algorithm presented above, we conducted experiments on several artificial data sets. We used 2-d and 3-d data spaces, crossing curves, non-crossing curves along which we generated data, different amounts of noise and different numbers of training examples. The results are promising, see Figs. 5 and 6 for an illustration

J.J. Verbeek et al. / Pattern Recognition Letters 23 (2002) 1009–1017

1015

Fig. 5. Results on a synthetic data set for our method (left) and GTM (right).

Fig. 6. Some of the PLs encountered (using 4, 6 and 12 segments, respectively), by the algorithm when fitting data distributed along a spiral.

(the thick segments in the plots are the fitted segments, the thin segments are inserted by the PL construction). Due to limited space we cannot present exhaustive experimental results but just some illustrations here. The good performance is due to (i) the flexibility of the method: the segments are fitted without being connected, and (ii) its incremental nature: segments are added one by one to avoid local optima and (iii) the ‘optimal’ number of segments is determined automatically. In Fig. 6 we illustrate, on the same data as in Fig. 1, several stages of the principal curve fitting. A difference between our method and other principal curve algorithms is the absence of a curvature penalty in our objective function (8). Although a curvature penalty is used in the construction of the curves, the algorithm is not too

sensitive to changes in its setting and it may even be omitted when modeling non-crossing curves. Regularization of the model complexity is instead introduced in (8) by (i) modeling the distribution of the latent variable by a uniform distribution and (ii) assuming a given level of noise all along the curve. This results in a regularization parameter r with a clear physical interpretation. The essence is that we avoid (i) distributions on the latent variable that have only finite support 4 and (ii) arbitrarily small noise levels. Experimentally we found that the applicability of the method is not limited

4

The distribution on the latent variable maximizing likelihood is a distribution with support only at the points on the curve where data project (Tibshirani, 1992).

1016

J.J. Verbeek et al. / Pattern Recognition Letters 23 (2002) 1009–1017

to cases where the naive assumption of a uniform distributed latent variable holds. Furthermore, it can be readily replaced by less rigid assumptions on the distribution of the latent variable, for example by a segment wise uniform distribution on the latent variable. Note that when one wants to fit other structures than curves, for example closed paths or branching structures, one only needs to replace the search procedure described in Section 3. In the case of closed paths the modifications needed are trivial since the same construction and optimization methods can be used. Our algorithm has a running time Oðkn2 Þ that is dominated by the insertion procedure, see Appendix A. For each of the k insertions the allocation takes Oðn2 Þ operations and hence takes Oðkn2 Þ in total. Constructing and optimizing all the PLs takes in total Oðk 3 Þ (taking the maximal number of iterations of 2-opt fixed). Computing the VRs in the optimization procedure after each segment allocation takes Oðnk 2 Þ operations in total. The total number of operations needed by the determination of the principal components is OðknÞ (taking the dimensionality of the data space fixed). Since in general k  n we arrive at a running time of Oðkn2 Þ. In applications involving large amounts of data the quadratic scaling with the amount of data may be problematic. In such cases more sophisticated and faster insertion methods may be used to allow for an Oðk 2 nÞ run time. For example only a small amount of candidate insertion locations may be considered using a method analog to the one used in (Verbeek et al., 2001). Some extensions on the presented method suggest themselves: Smooth curves. In some situations there may be a preference for smooth curves over PLs. In such cases one can use the found PL to assign a latent variable value to each datum (implying an ordering on the data). Subsequently, one may use some regression method to find a ‘good’ smooth curve. The found PL may also serve as an initialization of one of the other methods mentioned that find smooth curves. Obviate variance estimation. The use of an objective function that does not need to be tuned by the user would increase usefulness of the method.

We plan to use the minimum description length principle to find a parameter-free model selection criterion. Soft version. Currently we are investigating a ‘soft’ version of the presented method in which the ‘hard’ Voronoi partitions are replaced by a probabilistic weighting scheme. This way we hope to obtain better performance on very noisy data. Software implementing the k-segments algorithm in MA T L A B can be obtained from the homepage of the first author: http://www.science.uva. nl/jverbeek.

Appendix A. Efficient search for allocation of new segment At the start of the algorithm we compute the pairwise squared distances between all points in the data set, denote the resulting symmetric n  n matrix by D. In the optimization of the segments, we determine their VRs at each step. To do so we compute distances between all points and all segments, we store for each point xi the squared distance diVR to the closest segment. Let DVR ¼ ðd1VR ; d2VR ; . . . ; dnVR Þ> and VD ¼ ½DVR . . . DVR  be a square n  n matrix. If A is a matrix and a 2 R, then let M ¼ max ðA; aÞ denote the matrix M, where M i;j ¼ max ðAi;j ; aÞ. Let G ¼ maxðVD D; 0Þ. Then G i;j equals the decrease in squared distance for xi to its closest segment if we insert the zero-length segment at xj . Let 1 denote the 1  n matrix ½1 1 . . . 1. Then, arg maxf1Gg gives us the index of the datum maximizing the lower bound. References Bishop, C.M., Svensen, M., Williams, C.K.I., 1998. GTM: the generative topographic mapping. Neural Comput. 10, 215–234. Death, G., 1999. Principal curves: a new technique for indirect and direct gradient analysis. Ecology 80 (7), 2237–2253. Fritzke, B., 1994. Growing cell structures – a self-organizing network for unsupervised and supervised learning. Neural Networks 7 (9), 1441–1460. Gersho, A., Gray, R.M., 1992. Vector Quantization and Signal Compression. Kluwer Academic Publishers, Boston.

J.J. Verbeek et al. / Pattern Recognition Letters 23 (2002) 1009–1017 Hastie, T., Stuetzle, W., 1989. Principal curves. J. Am. Stat. Assoc. 84 (406), 502–516. Kegl, B., Krzyzak, A., Linder, T., Zeger, K., 2000. Learning and design of principal curves. IEEE Trans. Pattern Anal. Machine Intell. 22 (3), 281–297. Kohonen, T., 1995. Self-Organizing Maps. Springer, Berlin. Lawler, E.L., Lenstra, J.K., Rinnooy Kan, A.H.G., Shmoys, D. B. (Eds.), 1985. The Traveling Salesman Problem. Series in Discrete Mathematics and Optimization. Wiley, New York.

1017

Ripley, B.D., 1996. Pattern Recognition and Neural Networks. Cambridge University, Cambridge, UK. Tibshirani, R., 1992. Principal curves revisited. Stat. Comput. 2, 183–190. Verbeek, J.J., Vlassis, N., Kr€ ose, B., 2001. Efficient greedy learning of Gaussian mixtures, Technical Report IAS-UVA01-04, Computer Science Institute, University of Amsterdam, The Netherlands. Wilson, R.J., 1972. Introduction to Graph Theory. Oliver & Boyd, Edinburgh.